Cleaning Uncertain Time Series Based on Random Walk Sampling
Delin Zhou, Jizhou Sun, Bo Jiang, Toshboev Hudoyberdi, Jianchu Lin · 2022
In the era of big data, there are many different data types, and time series data makes up a large portion of observational data. However, when using sensors to collect time series data, the observation values are not very accurate due to the quality of the sensors and the observation errors. So how to estimate true data from multiple uncertain sensors' observations has becoming a hot research topic in IoT and big data fields. Several data cleaning methods for time series have been proposed, but many of them have strong constraints. Intuitively, if two observations are close, their contributions to true data estimation should be similar. According to this idea, in this paper we propose a novel data cleaning and truth discovery method named RWTD for time series based on random walk sampling. Distances between observations are calculated to build a state graph, and random walk sampling is used to estimate the weights of data sources. True data is estimated as the weighted sum of observations. Experimental results demonstrate the effectiveness of our method.